clear
set more off
* Load data set
use https://personal.lse.ac.uk/tenreyro/Pisch
* In the following estimation, ppml issues the message
* WARNING: The model appears to overfit some observations with y=0
* Note also that the covariance matrix is not computed and that no
* regressors are dropped
xi: ppml y i.iso_o
* The problems vanish if we estimate the same model using the
* equivalent syntax; notice that now 1 regressor is dropped
xi, prefix(_D) noomit i.iso_o
ppml y _D*
* The reason for the different behaviour is that with the first
* method we force _Iiso_1 to be the base category for iso dummies
* and that is a perfect predictor. With the second method we let
* ppml choose which variables are perfect predictors (just _Diso_1)
* and which one is the base category (_Diso_65)
* In short, it is generally preferable to generate all the dummies
* and include them all in the model, and let ppml choose what to do.